fluxgit-mcp-sidecar

mcp-name: io.github.fluxgit-hq/fluxgit-mcp-server

Safety-first Model Context Protocol (MCP) server for Git.

AI agents inspect. FluxGit keeps control.

A Rust MCP server with 38 contracts for AI code agents: 25 read-only tools and 13 human-gated operation tools (12 proposals plus cancellation of a pending proposal). The sidecar never executes a Git write; it bridges approved proposals to the FluxGit desktop application.

For context-budget comparisons, see the public Git context token benchmark. It publishes the fixture, raw outputs, scripts and checksums, including both the broad CLI sweep and a smaller hand-tuned counterexample.

Why this exists

AI coding agents are increasingly asked to navigate real repositories: explain branch state, summarize diffs, find lost commits, recommend safe next steps. To do this well, an agent needs Git context that is richer than git status and structured enough to reason over. To do this safely, an agent must never be able to silently mutate refs, force-push, discard work, or apply patches without a human approving the consequence.

Other MCP Git servers face a choice: stay strictly read-only (limited utility) or expose write tools directly (dangerous — agents hallucinate, prompts can be poisoned, mistakes are destructive). This sidecar chooses neither. Inspection is schema-validated and bounded; operations go through a write-with-UI-handshake: the agent proposes, FluxGit shows the preview, the user approves in the app, and FluxGit executes through its safety pipeline with restore points and audit.

What's exposed

25 read-only tools

13 write-with-UI-handshake tools

All 12 operation.preview.* proposals dispatch through the FluxGit gateway when configured. The sidecar POSTs the proposal, performs one bounded status read, and normally returns immediately with accepted: true, the canonical previewId, current status and nextAction.tool: "operation.status". The FluxGit app renders a “Requested by AI agent” approval card while human review continues asynchronously. Code 10003 is reserved for a bridge that is absent, invalid or unreachable; it is not a human-approval timeout.

All 12 preview schemas accept an optional bounded idempotencyKey. Reuse it only when retrying the same logical intent; the sidecar scopes it to the operation type so that retry resolves to the existing gateway proposal. Omit it for a new intent—even when the other arguments match—and the sidecar sends a fresh UUID-backed key. Preview tools therefore continue to advertise idempotentHint: false.

All write proposals require a free-text reason so the user sees the agent's justification in the approval modal. All reuse the same durable gateway lifecycle (pending → approved → executing → completed|failed, with rejection/cancellation/expiry branches) and the same Tauri bridge in the UI. Six shims cover pending, recoverable, approve, claim, reject and complete. After restart, Approved proposals may resume only after repo/ref revalidation and claim; Executing proposals are shown for explicit reconciliation and are never blindly re-executed. When an approved operation captures a restore point, the completion result exposes that recovery metadata so the agent can report it without guessing.

The boundary is deliberately fail-closed at approval time. FluxGit resolves the proposal's repoPath to its canonical open-repository id and requires it to match the repository the human is reviewing; an unresolved path, mismatch, or repository switch blocks execution. Tool arguments are validated before dispatch and again by the gateway. An optional declarative agent policy can deny proposals before a card opens. Its rules match the agentId the sidecar forwards (the sanitized, self-declared clientInfo.name, such as claude-code) and can limit that agent to operations, refs and, with pathConstraints, repository and worktree path prefixes. If FLUXGIT_MCP_AGENT_POLICY is configured but the file is missing, unreadable, malformed, or unsupported, the gateway does not start. With no configured policy, compatibility remains permissive, but per-operation human approval is still mandatory.

Other agents in the same repository (collision hints)

Every operation.preview.* result may carry an optional otherAgents object: the other agents agents.presence finds in the proposal's repository and, when both sides name their files (discard and commit paths, patch headers, plan steps, the submodule directory of submodulePointer), the overlappingPaths. It is informational only: it is added after the gateway decided, never blocks a proposal and never changes how it is approved. It is absent when agent detection is disabled.

"otherAgents": {
  "informational": true,
  "agents": [{ "agent": "codex", "state": "working", "sources": ["process", "session"],
               "branch": "codex/fix-login", "lastTool": "apply_patch",
               "files": ["src/login.rs"], "overlappingPaths": ["src/login.rs"],
               "pid": 4242, "lastActivityMs": 1790000000000, "mcpClient": null,
               "sameAgentAsCaller": false }],
  "agentCount": 1, "overlappingAgentCount": 1, "proposalPathsKnown": true,
  "omittedLikelyCaller": 1, "note": "Informational only: ..."
}

Detection is local and read-only: other agents' SQLite stores are opened read-only and only repository-relative paths, tool names, branches and times leave the detector. FLUXGIT_MCP_AGENT_DETECTION_DISABLED (any value) turns it off: agents.presence then returns 10011 and previews carry no hint.

Each sidecar also records which repositories its client used, so the desktop and other sidecars can see it: <run_dir>/presence/mcp/<pid>.json (0600) with the self-declared clientInfo name and version and, per repository, the canonical path, time and last tool name. Only calls that were accepted are recorded (a proposal the agent policy or a missing FluxGit refused is not); the file is removed on clean exit. FLUXGIT_MCP_PRESENCE_DISABLED turns it off.

Write protocol details

Every operation.preview.* call follows the same wire protocol. Example for operation.preview.merge:

1. Sidecar POSTs the proposal:

POST /v1/mcp/operation/preview/merge HTTP/1.1
Host: 127.0.0.1:59647
Content-Type: application/json

{
  "previewId": "1f3c5b9a-...-uuid",
  "agentId": "external-mcp-sidecar",
  "operationType": "merge",
  "repoPath": "/Users/dev/projects/checkout",
  "sourceRef": "feature/cart-redesign",
  "targetRef": "main",
  "reason": "Cart redesign work is complete; tests pass on the feature branch.",
  "strategy": "merge",
  "requestedAt": "2026-05-28T11:42:09.512Z"
}

2. Gateway responds 202 Accepted:

{ "previewId": "1f3c5b9a-...-uuid", "status": "pending", "expiresAt": "2026-05-28T11:47:09.512Z" }

3. Sidecar performs one bounded status read:

GET /v1/mcp/operation/status/1f3c5b9a-...-uuid HTTP/1.1
Host: 127.0.0.1:59647

If the proposal is still live, the preview tool returns promptly:

{
  "tool": "operation.preview.merge",
  "readOnly": false,
  "accepted": true,
  "previewId": "1f3c5b9a-...-uuid",
  "status": "pending",
  "nextAction": {
    "tool": "operation.status",
    "data": { "previewId": "1f3c5b9a-...-uuid" }
  }
}

This is a successful proposal submission, not a successful Git operation.

4. Client polls operation.status until the gateway reports a terminal state:

{
  "previewId": "1f3c5b9a-...-uuid",
  "operationType": "merge",
  "status": "completed",
  "result": {
    "commitSha": "9a8b7c6d...",
    "restorePointId": "rp_2026_05_28_1142",
    "conflicts": []
  }
}

completed returns isError: false. A live pending or approved proposal also returns as a successful accepted result, but it does not claim Git changed. Any non-completed terminal state (rejected, failed, expired, cancelled) returns isError: true with the structured payload, so the agent can report the real outcome instead of inventing one.

The same pattern applies to all 12 operation.preview.* tools. Only the request body fields and result shape differ; proposal submission, the one immediate read, asynchronous operation.status continuation and error semantics are shared. The public contract summary is maintained at fluxgit.com/features/mcp-agent-git.

Boundary: free shell vs FluxGit-powered

The sidecar speaks MCP without FluxGit installed. Standard Git inspection works (status, refs, history, reflog, diff.text, etc). The tools that require FluxGit return JSON-RPC error code 10001 with an upgradeHint pointing the agent at the install/configure flow.

Tier classification:

• Free shell — work with local git only: repo.brief, repo.scope, repo.status, repo.refs, repo.branchStack, repo.history, repo.reflog, commit.details, worktree.changes, worktree.list, submodule.status, diff.text, conflict.read, fleet.digest. agents.presence also needs no FluxGit: it reads local agent state (and FluxGit's MCP presence records when present).
• Hybrid — work locally with documented fallback, enriched by FluxGit: fleet.radar, diff.semantic, diff.semanticFallbacks, repo.conflictPreflight.
• FluxGit-required — return gateway_not_configured without FluxGit because synthesizing them from local refs alone would mislead the agent: safety.timeline, safety.eventDetails, flux.latestRestorePoint, flux.restorePoints, flux.restorePointDetails.
• Write handshake — route through FluxGit UI approval via the gateway handshake server. The 12 operation.preview.* tools return an accepted live proposal promptly and continue through operation.status; operation.cancel withdraws a pending proposal owned by the same agent. Code 10003 is used only when the bridge cannot accept or serve the handshake.

Quick start

Install the published crate (puts fluxgit-mcp-sidecar on your PATH):

cargo install fluxgit-mcp-sidecar --locked

To install the current source branch instead:

cargo install --git https://github.com/fluxgit-hq/fluxgit-mcp-server fluxgit-mcp-sidecar --locked

Or build from a clone:

# Build
cargo build --release

# Run as MCP server (stdin/stdout transport)
./target/release/fluxgit-mcp-sidecar

Connect any MCP-compatible agent

Paste the generic block below into any MCP host config. No client-specific install required.

{
  "mcpServers": {
    "fluxgit": {
      "type": "stdio",
      "command": "/absolute/path/to/fluxgit-mcp-sidecar",
      "env": {
        "FLUXGIT_MCP_HANDSHAKE_ADDR": "127.0.0.1:59647",
        "FLUXGIT_MCP_AUDIT_LOG": "/optional/path/to/audit.jsonl"
      }
    }
  }
}

FLUXGIT_MCP_HANDSHAKE_ADDR is the canonical bridge address generated by FluxGit Quick Connect. FLUXGIT_GATEWAY_ADDR and FLUXGIT_GATEWAY_URL remain compatibility fallbacks. The sidecar accepts only plain HTTP on a loopback host with an explicit port; a remote host, credentials, path, query, fragment, HTTPS, or missing port is rejected. Without a valid local bridge, the free-shell tier still works.

FLUXGIT_MCP_AUDIT_LOG enables an append-only JSONL audit log of every tools/call. Arguments are hashed; raw paths and identifiers are never written verbatim.

Semantic diff contract

diff.semantic is the most-used tool for AI agents and the easiest to misuse. The rule is strict:

A result may only be called semantic if data.supported is exactly true.

When the FluxGit semantic engine is not available (FluxGit app not running, gateway address not configured, or the repository not registered in FluxGit), diff.semantic returns:

{
  "tool": "diff.semantic",
  "readOnly": true,
  "data": {
    "supported": false,
    "fallback": "diff.text",
    "reason": "Semantic diff is not available in local sidecar fallback mode.",
    "textDiffArguments": { "repoPath": "...", "base": "...", "head": "...", "path": "..." }
  }
}

With the FluxGit app running and the repository registered in FluxGit, the same call is served by the FluxGit diff-engine through the gateway's read-only bridge and returns supported: true with per-file semantic hunks:

{
  "tool": "diff.semantic",
  "readOnly": true,
  "source": "fluxgit-gateway",
  "data": {
    "supported": true,
    "engine": "fluxgit-diff-engine",
    "files": [
      {
        "path": "src/main.rs",
        "fallbackToText": false,
        "hunks": [{
          "header": "fn main",
          "lines": [{
            "type": "modified", "oldLine": 3, "newLine": 3,
            "content": "let x = 2;", "oldContent": "let x = 1;",
            "changedTokens": ["2"], "oldChangedTokens": ["1"]
          }]
        }]
      },
      {
        "path": "logo.bin",
        "fallbackToText": true,
        "hunks": [],
        "reason": "The semantic engine could not parse this file (unsupported language, binary or unreadable source); use a text diff for it.",
        "textDiffArguments": { "repoPath": "...", "base": "...", "head": "...", "path": "logo.bin" }
      }
    ],
    "changedFiles": 2,
    "filesTruncated": false
  }
}

Honesty is per file, not just per call: files the engine could not parse arrive with fallbackToText: true, a reason, and ready-to-use textDiffArguments — never as synthesized semantic hunks. diff.semanticFallbacks follows the same split and, when connected, lists the engine's real per-file fallback records.

Connected agents must:

1. Call diff.semantic.
2. Read data.supported.
3. If true, use the semantic payload and label results as semantic — except entries with fallbackToText: true, which must be presented as text fallbacks.
4. If false, call diff.text with data.textDiffArguments and present results as a text-diff fallback.
5. Never infer function- or class-level moves from a text patch alone.

Allowed wording: "FluxGit reported a text-diff fallback for this file". Prohibited wording: "This is a semantic diff" when supported=false.

Presence file (which agent is working where)

Unless FLUXGIT_MCP_PRESENCE_DISABLED is set, each sidecar process keeps one local file, <FluxGit run dir>/presence/mcp/<pid>.json, so the FluxGit desktop app can show which coding agent is connected and which repository it is working on. It holds only the client name and version the agent declared in initialize, and per repository (at most 16) the checked repoPath, the time of the last call and the tool name. No other arguments, no results. Unlike the audit log, the repository path is stored as-is so the desktop can match it; the file is private (0600 in a 0700 directory), rewritten atomically, deleted on clean exit, swept after 24 hours, and writing it never affects a tool call. The client name is self-declared, not an authenticated identity.

Audit log

Unless FLUXGIT_MCP_AUDIT_DISABLED is set, the sidecar attempts to append each tools/call to the shared JSONL ledger. The gateway attempts human-decision appends through the same writer. FLUXGIT_MCP_AUDIT_LOG overrides the path; otherwise both processes use <FluxGit run dir>/audit/mcp.jsonl (including FLUXGIT_RUN_DIR):

{
  "id": "bdeca765-488c-4e2a-b86b-25cd734f2988",
  "timestamp": 1712345678901,
  "auditSchemaVersion": 1,
  "auditChainVersion": 1,
  "sequence": 42,
  "segmentId": "7ab6fa7a-c5bf-4d82-86a8-26b4728b5acd",
  "previousHash": "sha256:...",
  "entryHash": "sha256:...",
  "tool": "repo.status",
  "event_type": "tool_call",
  "repo_scope": "repoPath:sha256:...",
  "args_fingerprint": "sha256:...",
  "risk": "read",
  "approval": "none",
  "result": "success",
  "session_id": "my-agent",
  "duration_ms": 12,
  "summary": "...",
  "readOnly": true,
  "sidecarReadOnly": true,
  "signature": "base64url-ed25519",
  "signatureKeyId": "1a2b3c4d5e6f7a8b",
  "signatureVersion": 3
}

Sensitive paths and identifiers are hashed, never stored verbatim. One stable cross-process lock protects validation, rotation and the complete append plus sync. Lines are limited to 256 KiB. The active segment rotates at 4 MiB and at most four rotated segments are retained, with a signed checkpoint when signing is configured.

Per-entry Ed25519 signatures (shipped 2026-05-28)

Audit signing is opt-in. When FLUXGIT_MCP_AUDIT_SIGN_KEY points to a PEM PKCS8 Ed25519 private key, every appended entry is signed with that key. Signed entries add:

• signature — base64url (no padding) Ed25519 signature over the canonical JSON of the entry without the signature field.
• signatureKeyId — 16-char hex prefix of the matching public key, so rotated keys can co-exist in the same JSONL.
• signatureVersion: 3 — current chained signing domain; the key id, sequence, previous hash and entry hash are included in the signed bytes.

Canonical JSON rule (verifier must match exactly): recursively sort every object's keys lexicographically by UTF-8 byte order; arrays preserve order; strip signature; serialize compactly. For versions 2 and 3, keep signatureKeyId and signatureVersion in the signed object. For legacy signed entries with no version, the verifier also strips signatureKeyId.

If the env var is unset, new entries are chained but unsigned for backward compatibility. If it is explicitly set, an empty, missing, unsafe, oversized or invalid key fails audit startup closed; it never degrades to unsigned output.

Verifying an audit log

The sidecar binary doubles as a verifier:

fluxgit-mcp-sidecar verify-audit /path/to/mcp.jsonl --pubkey /path/to/install.pub.pem

The CLI streams the active file and retained rotations with bounded memory. It validates sequence, hashes, segment names, checkpoints and signatures, then reports only bounded counters (entries, chained, legacy, signed, unsigned, segments and the retained sequence range). Legacy per-entry records remain readable and verifiable but are reported as legacy, never as part of the tamper-evident chain. For a strict evidence gate, run:

fluxgit-mcp-sidecar verify-audit /path/to/mcp.jsonl --pubkey /path/to/install.pub.pem --require-signed

Exit code is 0 on success, 3 for malformed data, broken chain/rotation, a bad signature and, in strict mode, any unsigned entry; usage errors return 2.

Programmatic full-ledger verification uses verify_audit_ledger; the older verify_audit_event_signature remains available for compatible per-entry checks. A local chain cannot prove deletion or replacement of the entire retained history (or its local checkpoint) without an independently trusted external anchor. Signed retained entries do prevent an attacker without the private key from recomputing a modified chain.

Audit configuration (including an explicit signing key) fails closed at startup. A later filesystem/full-disk append failure is logged as degraded but does not undo a tool response or an already-durable gateway lifecycle transition; the gateway lifecycle journal remains authoritative for recovery.

Protocol details

The server supports two protocol eras:

• 2026-07-28 (preferred, stateless): call server/discover, then include params._meta.io.modelcontextprotocol/protocolVersion and params._meta.io.modelcontextprotocol/clientCapabilities on every request. Modern results carry resultType: "complete" and server metadata; list results add ttlMs and cacheScope. tools/list includes title, inputSchema, outputSchema and annotations. tools/call includes both presentational content and the same payload in structuredContent. Only the proposal tools and operation.status advertise previewId, status, accepted and nextAction in their outputSchema; schema objects are open unless they say "additionalProperties": false.
• 2024-11-05 (legacy compatibility): older hosts continue to use initialize. Modern-only fields are omitted from legacy results.

Stdio output is standard newline-delimited JSON-RPC 2.0: exactly one compact JSON value per line. The JSON in a result's content[0].text is compact too (no indentation), with the same keys and values as structuredContent. Pre-standard Content-Length framing remains accepted as input only for old FluxGit clients; the server never emits it. Frames are limited to 8 MiB. JSON-RPC notifications receive no response, and request methods sent without an id are not executed.

There are 38 tools in modern tools/list: 25 advertise annotations.readOnlyHint: true; the 12 operation.preview.* tools and operation.cancel advertise readOnlyHint: false. These annotations describe effects for the host; they are not authorization.

Error codes:

Status

This is a working MCP server. The read-only surface and all 12 operation.preview.* routes are implemented; operation.cancel manages only a pending proposal owned by the same self-reported agent id. The write handshake renders an approval card in FluxGit and completes through the app's guarded pipeline. Clients receive structured lifecycle results instead of a synthetic success. clientInfo.name is sanitized attribution for policy, quota and audit; it is self-reported and must never be treated as authenticated identity.

Roadmap

• End-to-end demo video — public recording of the agent-proposes → user-approves → FluxGit-executes loop, captured from a live install.
• Audit log exportable CSV/JSON — shipped: per-entry Ed25519 signing (2026-05-28). Remaining: exportable CSV/JSON and retention policy for the FluxGit app's audit panel.
• HTTP / SSE transport — for cloud / shared MCP host deployments.
• Official MCP Registry — io.github.fluxgit-hq/fluxgit-mcp-server is live and resolves to the published fluxgit-mcp-sidecar crate.

License

Apache-2.0. See LICENSE.

Related

• FluxGit — the desktop app that produces the FluxGit-powered context.
• MCP agent Git — public product and protocol overview.
• Claude Code, Cursor, and Codex — setup and workflow guides.
• Git context token benchmark — reproducible fixture, raw outputs, scripts and checksums.
• Public source — this Apache-2.0 server.